Real-Time Computer Vision for Transportation Safety using Deep Learning and IoT

With the advancements in technology and application orientations, the field of computer vision is evolving at a rapid rate. Real-Time systems capable of detecting objects with high precision has been a spotlighted field of research. Autonomous traffic systems, self-driving cars and advanced driver assistance systems (ADAS) are the areas where computer vision plays an extensive role. This paper employs YOLO v4 deep learning based computer vision system for detecting vehicles, animals and pedestrians in real-time with high precision scores. The system is based on cross stage partial (CSPDarknet53) convolutional neural network serving as the backbone network for object detection. For precise detection of the entities, 16,300 image sequences were utilized for training the algorithm. For fast computation and detection, the algorithm was deployed on NVIDIA GTX 1060 GPU yielding into effective frames per second processing capabilities. The algorithm was trained for cars, bikes, pedestrians, cows and dogs yielding precision scores of 0.9912, 0.9868, 0.9783, 0.9770 and 0.9726 respectively. Also, the system alerts the vehicle in case of pothole areas nearby to ensure safety. Survey-based pothole and animal hotspots were collected and stored in the server geo-database. The vehicle approaching the hotspot triggers the alert system with acoustic feedback. The deployed real-time computer vision system serves as an effective system for real-time transportation entities detection applications laying a solid foundation for intelligent transportation systems, autonomous vehicles and advanced traffic applications.